Earthly Machine Learning

Citation: Dueben, P., Bauer, P., Fuhrer, O., Koldunov, N., & Kristiansen, J. (2026). Machine learning is revolutionizing weather forecasting – the next step is a change in how we work. arXiv preprint arXiv:2606.25076v1.

Key Takeaways:

• A Shift in the Forecasting Value Chain: While machine learning has rapidly achieved competitive skill in weather predictions, the next critical phase is a complete evolution of working practices and operating models. This transition will fundamentally reshape how models are coded, how observational data is exploited, and how forecasts are verified and turned into public services.
• The Rise of Agentic AI and Automated Workflows: The traditional, slow-paced manual approach to Earth-system model development is giving way to AI-assisted and agentic workflows. Large Language Models (LLMs) and software agents will increasingly handle the writing, testing, optimizing, and porting of code, shifting the role of human scientists from active programmers to supervisors, test designers, and monitors.
• Modernized Software and Open Data Stewardship: To benefit from rapid AI innovation, meteorological centres must adopt industry-standard software frameworks (like Python, PyTorch, and JAX) and modular code structures. This runs parallel to a major shift in data stewardship, moving toward highly compressed, cloud-native, open-access datasets that can be efficiently searched and streamed by both humans and AI agents.
• On-the-Fly Generative Emulation: Emerging generative machine learning foundation models will enable interactive, real-time "what-if" simulations and climate scenario exploration. Instead of moving or storing massive datasets, models can recreate specific atmospheric states on demand, though this introduces a critical need for rigorous verification techniques to distinguish physical realism from AI hallucinations.
• Organizational Risks and Preserving Expertise: Adapting to mixed human-AI environments presents risks, including the potential erosion of expert scientific knowledge if critical workflows are fully offloaded to AI. Weather and climate centres must proactively design transparent model diagnostics, continuous testing, and educational practices to maintain human operational understanding.

Citation:

Rampal, N., González-Abad, J., Addison, H., Baño-Medina, J., Bettolli, M. L., Blasone, V., Booth, B., Coppola, E., Di Gioia, S., Oldham-Dorrington, J., Doury, A., Engelbrecht, F., Fuentes-Franco, R., Gibson, P. B., Glawion, L., Hardy, C., Ivanov, M., Lee, H. K., Legasa, M. N., Olmo, M., Orr, A., Polz, J., Rogers, M. S. J., Schillinger, M., Sharma, S., Soares, P. M. M., Sobolowski, S., Steinkopf, J., Tang, W., Tian, J.-B., Tomé, R., Wang, K.-C., Wang, Y.-C., Watson, P. A. G., Wetherell, T., Widmann, M., & Gutiérrez, J. M. (2026). CORDEX-ML-Bench: A Benchmark for Data-Driven Regional Climate Downscaling—Experiment Design and Overview. WCRP-CORDEX Machine Learning Task Team.

Key Takeaways:

• Establishing a Standardized Global Benchmark: The paper introduces CORDEX-ML-Bench, the first coordinated multi-domain, multi-architecture benchmarking framework explicitly designed to standardize machine learning (ML) models for regional climate downscaling. The framework enables researchers to evaluate and compare models consistently using open-source datasets and metrics, focusing on a 20× spatial resolution increase (from ~200 km down to ~10 km grids) for daily precipitation and maximum temperature. The initial benchmark targets three highly diverse geographical regions: the European Alps, New Zealand, and Southern Africa.
• The Critical Extrapolation Gap of Historical Training: A foundational finding of the benchmark is that ML models trained strictly on historical climate data systematically underestimate future climate change signals, including extreme warming and intensive precipitation. This reveals a major vulnerability in traditional historical-only statistical downscaling methods and proves that incorporating future climate projections into training datasets (the emulator approach) is vital for producing physically credible long-term projections.
• Generative AI Outperforms for Precipitation Extremes: In an evaluation of 40 independently developed ML configurations, generative approaches (such as diffusion models, flow matching, and Generative Adversarial Networks) consistently outperform deterministic regression models at downscaling precipitation. They excel at capturing highly localized spatial variability and heavy-tailed extreme events, whereas deterministic models suffer from spatial "oversmoothing". However, deterministic architectures remain highly competitive for predicting daily maximum temperature.
• Balancing Accuracy and Computational Cost: While highly complex diffusion models (like the top-ranked RCMGEM-mv-orog) achieve outstanding accuracy, they are computationally intensive. In contrast, flow-matching and GAN-based models achieve a highly favorable skill-to-compute ratio, lowering inference costs by one to two orders of magnitude. This makes them highly practical solutions for generating large, multi-model ensemble climate projections under restricted computational budgets.

Citation: Henn, B., Bretherton, C. S., Kodunov, N., Lessig, C., Molina, M. J., Arcomano, T., Watt-Meyer, O., Couairon, G., Singh, R., Brunstein, R., Hasson, Y., Jost, A., Brenowitz, N., Manshausen, P., Cresswell-Clay, N., Durran, D., Hall, K. J. C., Yuval, J., Kochkov, D., Hoyer, S., & Lopez-Gomez, I. (2026). AIMIP Phase 1: systematic evaluations of AI weather and climate models. arXiv preprint.

• A New Benchmarking Era for AI Climate Models: AIMIP Phase 1 establishes the first systematic intercomparison framework for artificial intelligence weather and climate models (AIWCMs). It defines a common experimental protocol, standardizes CMIP-compatible output formats, and provides an open dataset to evaluate how different AI architectures influence long-term climate simulation behaviors.
• Standardized Historical Simulation Protocol: Under the Phase 1 protocol, participating models are trained exclusively on historical ERA5 atmospheric reanalysis data from 1979 to 2014 and run through a 10-year out-of-sample test period (2015–2024). To prevent overfitting, models are forced only by specified sea surface temperatures (SST) and sea ice concentrations (SIC), with direct greenhouse gas inputs (like CO2 concentrations) strictly excluded.
• Excellent Representation of Baseline Climate and ENSO: The initial evaluations of the eight participating AI models demonstrate that they represent time-mean climate averages and natural variability patterns—such as the El Niño-Southern Oscillation (ENSO)—just as well as, or in some cases with lower systematic biases than, conventional physically-based climate models like the NOAA GFDL-CM4.
• The Out-of-Sample Warming Gap: A primary weakness identified across several AI models is their struggle to accurately replicate global warming trends during the out-of-sample test period (2015–2024). Because greenhouse gases like CO2 are omitted as direct predictors to avoid overfitting, some models fail to translate rising ocean temperatures into the full magnitude of observed atmospheric warming.
• Extreme Extrapolation Remains a Challenge: When subjected to extreme, highly out-of-sample sensitivity experiments where sea surface temperatures are uniformly raised by +2 K and +4 K, the AI models diverge significantly. They produce highly inconsistent and sometimes physically implausible responses (such as simulated cooling over land), highlighting that projecting unseen future climates remains a key development challenge for the AI climate modeling community.

Citation: Glaser, Y., Stopa, J. E., Wolniewicz, L. M., Foster, R., Vandemark, D., Mouche, A., Chapron, B., & Sadowski, P. (2025). WV-Net: A Foundation Model for SAR Ocean Satellite Imagery. Artificial Intelligence for the Earth Systems, e250003. DOI: 10.1175/AIES-D-25-0003.1

Key Takeaways

First Foundation Model for Open-Ocean SAR Imagery: WV-Net represents the first-ever foundation model designed specifically for open-ocean sea surface images, utilizing a massive dataset of nearly 10 million unannotated C-band synthetic aperture radar (SAR) wave mode images collected globally by the Sentinel-1 satellite mission.

Overcoming the Annotation Bottleneck: By leveraging contrastive self-supervised learning (SimCLR), the model learns highly robust, general-purpose representations of complex geophysical signatures directly from raw, unlabeled imagery—bypassing the traditional bottleneck of expensive manual expert annotation.

Outperforming General-Purpose Computer Vision Models: The model's specialized ocean-domain embeddings consistently beat standard models pretrained on natural images (like ImageNet) across key downstream tasks, including estimating wave height, predicting air-sea temperature differences, and identifying 12 distinct atmospheric and oceanic phenomena.

High Data Efficiency & Fine-Tuning Stability: WV-Net scales exceptionally well in data-constrained settings, delivering strong performance with as few as 100 labeled training examples. Additionally, it exhibits greater robustness to hyperparameter selections during fine-tuning, dramatically reducing the need for broad, computationally heavy optimization sweeps.

Optimizing Domain-Specific Augmentations: The researchers discovered that standard computer vision augmentations adapted for radar data (such as mixup, color inversions, rotations, and sharpness adjustments) were crucial to bridging the domain gap, while complex, domain-specific signal filtering (such as random notch filtering) actually degraded model performance.

Citation: Ji, C., Mu, M., Qin, B., Lian, T., Yuan, S., Feng, J., Song, S., Wei, Y., Dai, G., Wang, J., & Fang, X. (2025). Toward skillful forecasting of super El Niño events using a diffusion-based westerly wind burst parameterization. npj Climate and Atmospheric Science (Published in partnership with CECCR at King Abdulaziz University). https://doi.org/10.1038/s41612-025-01158-x

Key Takeaways:

• Innovative Generative AI Parameterization: The study introduces a state-of-the-art Denoising Diffusion Probabilistic Model (DDPM) to parameterize westerly wind bursts (WWBs). These wind bursts are critical, episodic atmospheric events that inject wind energy into the Pacific, playing a pivotal role in triggering super El Niños. This new generative AI framework successfully captures the complex, joint modulation of wind bursts by both slow-varying oceanic states and rapid atmospheric processes.
• Superior Representation of Wind Burst Physics: Traditional schemes rely heavily on ocean-state indicators like the warm pool eastern edge, which fails to capture high-frequency atmospheric noise. By incorporating multiple physical conditions—Sea Surface Temperature Anomalies (SSTA), Outgoing Longwave Radiation Anomalies (OLRA), and Sea Level Pressure Anomalies (SLPA)—the DDPM-based scheme dramatically improves the simulated frequency, intensity, duration, and spatial distribution of wind bursts compared to observational data.
• Drastic Improvements in Super El Niño Intensity Predictions: When coupled online with the Community Earth System Model (CESM), the DDPM scheme significantly outperforms both standard control runs and traditional parameterization schemes. It accurately predicts the absolute amplitude of historic super El Niño events—specifically the 1982/83, 1997/98, and 2015/16 events—by correcting the severe underestimations found in baseline climate models.
• Mitigation of Seasonal Phase-Locking Bias: A persistent challenge in climate modeling is "seasonal phase-locking" prediction bias, where models incorrectly project a double-peak warming cycle (peaking in summer, weakening, then re-intensifying in winter). The DDPM scheme overcomes this issue by generating stronger and more realistically eastward-shifted wind stress anomalies, which correctly trigger the positive dynamical feedbacks (such as the Bjerknes feedback) necessary to sustain a steady, natural warming progression toward a single December peak.

Citation: Kaack, L. H., Donti, P. L., Strubell, E., Kamiya, G., Creutzig, F., & Rolnick, D. (2022). Aligning artificial intelligence with climate change mitigation. Nature Climate Change, 12, 518–527. https://doi.org/10.1038/s41558-022-01377-7

Main Takeaways:

• Three Layers of AI's Climate Footprint: The authors propose a framework that splits machine learning's climate impact into three distinct categories — the energy and hardware emissions of computing itself, the immediate effects of specific ML applications, and the broader system-level changes that ML induces across society. The categories that are easiest to measure (like the electricity used to train a model) are likely not the ones with the largest effects, which is why most current discussions of "AI and climate" capture only a sliver of the real picture.
• Computing Is a Small Slice — For Now: The entire global ICT sector accounts for roughly 1.4% of global greenhouse gas emissions, and AI workloads are only a fraction of that. But the trajectory is steep: at Facebook, ML training compute has been growing about 150% per year and inference compute about 105% per year, far outpacing efficiency gains. Even striking efficiency wins — like Google's TPU being 30–80 times more energy-efficient than contemporary CPUs or GPUs — can be swamped by raw growth in demand.
• The "Internet of Cows" Problem: ML is a general-purpose tool, which means it's just as good at accelerating oil and gas exploration or scaling up cattle farming (an industry already responsible for about 9% of global emissions) as it is at forecasting solar power or optimizing data center cooling. Whether AI is net-positive or net-negative for the climate is genuinely undetermined, and depends on which applications get funded, deployed, and regulated.
• System-Level Effects May Dwarf Everything Else: The largest climate impacts of AI may come not from training runs or even individual applications, but from how ML reshapes society — through rebound effects (efficiency gains that drive more consumption), technological lock-in (autonomous cars entrenching private vehicle travel over transit and rail), and ML-powered recommender systems that boost demand for emissions-intensive goods. These effects are the hardest to quantify but potentially the most consequential, and the authors argue they need to be built into climate scenario modeling — something the IEA, EIA, and IPCC's Shared Socioeconomic Pathways largely don't do today.

Machine learning for the physics of climate

Citation: Bracco, A., Brajard, J., Dijkstra, H. A., Hassanzadeh, P., Lessig, C., & Monteleoni, C. (2025). Machine learning for the physics of climate. Nature Reviews Physics, 7, 6–20. https://doi.org/10.1038/s42254-024-00776-3

Main Takeaways:

• Breaking the El Niño Spring Barrier: For decades, forecasts of the El Niño Southern Oscillation hit a hard wall at roughly 6 months lead time — a limit known as the spring predictability barrier. Convolutional neural networks trained on a mix of climate model and reanalysis data have shattered this ceiling, delivering skillful forecasts at 17 months out, with newer architectures pushing to 21–24 months. ML models can also now anticipate which type of El Niño will develop (eastern vs. central Pacific), which matters enormously because the two flavors produce very different regional impacts around the world.
• Weather Forecasting at a Fraction of the Cost: A new generation of ML weather emulators — Pangu-Weather, GraphCast, FourCastNet, FuXi, NeuralGCM — now match or beat the European Centre's flagship physics-based forecasting system on most variables, including hurricane tracks, while running orders of magnitude faster. They achieve this with surprisingly compressed state representations: roughly 10 vertical atmospheric levels and 0.25° horizontal resolution, compared to 100+ levels and 0.1° in conventional models. The catch is that these models can violate basic physics — geostrophic balance, energy conservation, the butterfly effect — which currently blocks naive extension to climate timescales.
• Hybrid Models Are Eating the Climate Stack: Pure ML works for short-range forecasts, but for climate-length runs the field is converging on hybrid architectures that pair a traditional dynamical core with neural-network parameterizations of sub-grid processes like clouds, turbulence, and gravity waves. Google's NeuralGCM exemplifies the approach and already reduces biases in tropical cyclone frequency and tracks. A telling case study on the quasi-biennial oscillation showed that an offline-trained neural network produced unstable, unphysical results — but retraining just two layers online, coupled to the model, recovered the correct physics. Offline-only or online-only training each fail in characteristic ways; the mix is what works.
• The Data Wall Is the Real Bottleneck: Climate ML has less than 50 years of dense satellite-era observations to work with, and those observations are heavily biased toward the atmosphere and ocean surface — a single, spatiotemporally correlated realization of one climate. This limits how confidently ML models can extrapolate to warmer, unseen climates, which is exactly what climate projection requires. The path forward involves three parallel bets: hybrid physics-ML models that bake in conservation laws, large-scale "foundation models" for weather and climate trained across simulations and observations together (efforts like ClimaX and AtmoRep are early examples), and rare-event sampling strategies to handle the extremes that matter most for adaptation policy but are by definition underrepresented in any training set.

Citation: Benson, V., Bastos, A., Reimers, C., Winkler, A. J., Yang, F., & Reichstein, M. (2025). Atmospheric transport modeling of CO2 with neural networks. Journal of Advances in Modeling Earth Systems, 17, e2024MS004655. https://doi.org/10.1029/2024MS004655

Main Takeaways:

• A New Benchmark for AI Carbon Tracking: The authors introduce CarbonBench, the first systematic benchmark dataset designed specifically for training and evaluating machine learning emulators of Eulerian atmospheric transport. Built from CarbonTracker CT2022 inversions and ObsPack station observations, it ships at three resolutions (the coarsest being 5.625° × 10 vertical levels × 6h) and is engineered to plug directly into modern deep learning pipelines — opening atmospheric carbon modeling to the broader ML community.
• SwinTransformer Wins, Decisively: Of the four architectures tested (UNet, GraphCast, SFNO, and SwinTransformer), the SwinTransformer reaches near-perfect emulation with a 90-day R² above 0.99 and stays stable in physically plausible forward runs for over three years — a regime where neural PDE solvers typically blow up. At measurement stations, it actually captures the seasonal cycle in Svalbard better than TM5, the conventional model it was trained to emulate, possibly due to differences in boundary layer transport near the poles.
• Physics Tricks Were the Unlock: Out of the box, the neural networks were unstable — especially the mesh-based UNet and GraphCast. Two simple physics-aware adjustments fixed this across all four architectures: centering the CO2 input field at each timestep to remove the covariate shift from steadily rising atmospheric CO2 (called CentFlux), and a post-hoc mass fixer that rescales predicted mass to match the surface flux budget. The result is mass conservation with RMSE of just 0.00058 PgC against a total atmospheric carbon mass of ~865 PgC — effectively negligible.
• Speed Isn't the Selling Point (Yet): Unlike AI weather models, which famously outpace numerical forecasting by orders of magnitude, the SwinTransformer is not significantly faster than TM5 at this resolution — about 1.5 seconds for a 30-day run on an A40 GPU versus a few minutes for TM5 on 24 CPUs. The real promise lies elsewhere: the networks are fully differentiable (useful for inverse modeling of surface fluxes), natively support batched ensembles, and scale better to high resolution where conventional solvers become prohibitively expensive — exactly the regime where current CO2 inversions struggle most.

Citation: Zhu, X. X., Xiong, Z., Wang, Y., Stewart, A. J., Heidler, K., Wang, Y., Yuan, Z., Dujardin, T., Xu, Q., & Shi, Y. (2026). On the foundations of Earth foundation models. Communications Earth & Environment, 7, 103. https://doi.org/10.1038/s43247-025-03127-x

Main Takeaways:

• Current Earth AI Models Are Missing the Point: Researchers have identified eleven features that an ideal Earth foundation model must have — including geolocation awareness, multi-sensor integration, physical consistency, and carbon minimization — yet no existing model comes close to checking all eleven boxes. Most models focus on only one or two features, leaving a major gap between what we have and what we actually need to tackle real-world climate and environmental challenges.

• The Data Situation Is More Lopsided Than You'd Think: There are now over 1,000 active remote sensing satellites generating nearly 100 petabytes of open satellite data — but labeled datasets used to train AI models account for less than 0.1% of that archive. This massive imbalance is precisely why self-supervised foundation models, which can learn from unlabeled data, are so critical for Earth science going forward.

• Weather AI Is Already Dramatically More Efficient — But Incomplete: Models like FourCastNet can generate a week-long global weather forecast in under two seconds on a single GPU, using roughly 12,000 times less energy than traditional forecasting systems. Despite this leap in efficiency, major gaps remain: models struggle beyond two-week forecasts, long-term climate projections drift due to incomplete energy balance, and connecting fine-scale satellite imagery with coarse climate models remains largely unsolved.

• What Comes After the Ideal Model: Once a true Earth foundation model exists, the authors argue the most exciting frontier is using it to build an "Earth Embedding" — a compact, unified representation of our entire planet that researchers worldwide could query without ever touching raw satellite data. Beyond that, challenges like machine unlearning (making models forget sensitive imagery), adversarial defenses, and continual learning as the climate itself changes will define the next generation of Earth AI research.

Citation: Olivetti, L., & Messori, G. (2025). Whose weather is it? A fairness framework for data-driven weather forecasting. Environmental Research Letters, 20, 121006. https://doi.org/10.1088/1748-9326/ae21f5

Main Takeaways:

• AI Weather Models Aren't Fair to Everyone: The latest generation of AI-powered weather forecasts improves predictions globally — but not equally. Using ECMWF's AIFS model as a case study, the authors show that wealthier and more densely populated areas consistently receive a higher share of forecast improvements compared to poorer and more rural regions, violating basic fairness criteria borrowed from the algorithmic fairness literature.

• Two Measurable Fairness Tests — Both Failed: The paper proposes two concrete criteria: statistical parity(improvement rates should be similar across income groups) and conditional independence (a region's GDP or population density should not predict whether it benefits from the new model). Across nearly all tested variables and forecast lead times, AIFS fails both tests at the 0.01 significance level — meaning the disparity is not a statistical fluke.

• Extreme Weather Is Where the Gap Hurts Most: For standard temperature and wind forecasts, gaps between rich and poor regions are modest. But for cold extremes, the fairness gap is especially pronounced — precisely the events where accurate early warnings matter most for vulnerable populations with fewer resources to adapt.

• Fixing It Is Technically Feasible: Unlike traditional physics-based models, AI weather models offer genuine design levers for fairness. The authors describe two practical approaches: adding penalty terms to the loss function (such as the Hilbert–Schmidt Independence Criterion) to reduce associations with protected variables, and using geographically adaptive weighting that iteratively compensates for emerging performance gaps — without necessarily sacrificing global accuracy.